GPAI Code of Practice: the Commission's assessment and the Article 53 evidence it expects

What the obligation asks for

Art. 53(1) of Regulation (EU) 2024/1689 places four duties on the provider of a general-purpose AI model: technical documentation with at least the Annex XI content, kept up to date and given to the AI Office and national competent authorities on request (point (a)); information and documentation with at least the Annex XII content for providers of AI systems who integrate the model (point (b)); a policy to comply with Union copyright law that identifies rights reservations under Article 4(3) of Directive (EU) 2019/790 (point (c)); and a public summary of the training content in the AI Office template (point (d)). Art. 53(2) lifts points (a) and (b) for a model released under a free and open-source licence with public parameters, weights, architecture and usage information, and lifts nothing for a model with systemic risk; Art. 53(3) requires the provider to cooperate with the Commission and national competent authorities. Art. 55(1) adds four duties for a model with systemic risk: evaluation with documented adversarial testing (point (a)), assessment and mitigation of systemic risks at Union level (point (b)), serious incident tracking and reporting (point (c)) and cybersecurity for the model and its physical infrastructure (point (d)). The Art. 53 obligation record lists the five testable elements this guide covers: Annex XI completeness, Annex XII completeness, cross-document consistency, evaluation result traceability and training-content summary structure. The duties apply from the GPAI application date 2025-08-02In force, the Commission's powers to request information and to fine under Art. 91 and Art. 101 apply from the enforcement date 2026-08-02In force, and Regulation (EU) 2026/1744 amended neither article.

Art. 56(6) requires the AI Office and the Board to assess whether a code of practice covers the Art. 53 and Art. 55 obligations and to publish their assessment of its adequacy. The Commission published that assessment as an opinion, with the Board's opinion alongside, stating that the GPAI Code of PracticePublished is an adequate voluntary tool for providers of general-purpose AI models to demonstrate compliance with the AI Act; the signatory taskforce document on the same Commission page records the Code as assessed as adequate pursuant to Art. 56(6). The assessment fixes the route; the duties stay in the articles. Art. 53(4) and Art. 55(2) let a provider rely on a code of practice under Art. 56 to demonstrate that the Art. 53(1) and Art. 55(1) duties are met until a harmonised standard is published, and none is cited in the Official Journal as of the anchor date; a provider that neither adheres to the Code nor applies a harmonised standard, of which none is cited, demonstrates alternative adequate means for assessment by the Commission. A presumption of conformity attaches only to a European harmonised standard, and no harmonised standard for general-purpose AI models is cited in the Official Journal as of the anchor date on the standards page, so no presumption is available on either route. The Code says the same of itself: Objective A of each chapter states that adherence does not constitute conclusive evidence of compliance with the obligations under the AI Act. The Code's three chapters divide the evidence as follows. The Transparency chapter, Commitment 1, covers Art. 53(1)(a) and (b) and Annex XI Section 1 and Annex XII through the Model Documentation Form (Measure 1.1), the provision of information to the AI Office under Art. 91 and to downstream providers (Measure 1.2), and quality, integrity and security of the documented information (Measure 1.3). The Copyright chapter, Commitment 1, covers Art. 53(1)(c) through Measures 1.1 to 1.5. The Safety and Security chapter covers Art. 55(1) through the Safety and Security Framework (Commitment 1), systemic risk identification, analysis, acceptance and mitigation (Commitments 2 to 6), the Safety and Security Model Report (Commitment 7), serious incident reporting (Commitment 9) and, under Measure 10.1, the Annex XI Section 2 documentation that Art. 53(1)(a) adds for a model with systemic risk. Art. 53(1)(d) sits in no chapter: the Commission's explanatory notice and template for the public summary of training content is a separate instrument, and its accompanying notice states that use of the template is mandatory while adherence to the Code is voluntary.

Inputs

The protocol page for PR-11-DOC v1.0 lists every input. The general-purpose AI model route needs the following from that list, plus one input from PR-15-ROB v1.0.

• The provider determination record and the open-source exception assessment record from PR-CLS v1.0, and, where the Art. 51 gate is in question, the training compute determination record.
• The Annex XI documentation and the Annex XII information pack as they stand, with version identifiers; where the provider uses the Model Documentation Form, the completed form with its document version number and last-updated field.
• The public model card and the published training-content summary, each with its URL.
• The copyright policy as a single document, with the public summary where one is published, the point of contact and the complaints mechanism named in Measure 1.5 of the Copyright chapter.
• The record of whether the provider has signed the Code, which chapters, and, for a non-signatory, the alternative adequate means it presents under Art. 53(4) and Art. 55(2).
• For a model with systemic risk: the Safety and Security Framework, the current Model Report, the serious incident register and the model evaluation report and adversarial testing campaign record from PR-15-ROB v1.0.
• Write access to a run directory for the frozen input set, the evidence index, the gap tables and the report.

Procedure

Steps 1 to 12 below are the general-purpose AI model subset of PR-11-DOC, renumbered for this guide; the cite line on each step names the protocol step and the clause it derives from. Every model name, hash, identifier, path and number in the code blocks is a synthetic example, and none is a requirement of the Act.

1. Record the documentation route as Annex XI and Annex XII for the model version under review.

   Record Route determination record: Annex XI and XII

   [Regulation (EU) 2024/1689, Art. 53(1); PR-11-DOC step 2]

2. Take the provider determination, the open-source exception and the training compute records from PR-CLS.

   Record Art. 53 scope record with the two PR-CLS record identifiers and the systemic-risk gate

   [Regulation (EU) 2024/1689, Art. 53(1) and Art. 53(2); PR-11-DOC step 4]

3. Record the means of demonstration: the chapters signed, or the alternative adequate means the provider presents.

   Record Code of practice adherence statement or alternative-means record (Art. 53(4))

   [Regulation (EU) 2024/1689, Art. 53(4) and Art. 55(2); GPAI Code of Practice, Objective A of each chapter]

4. Build one row per Annex XI and XII point, mapped to its Model Documentation Form field and audience marker.

   Record Annex XI and XII checklist template with the Model Documentation Form field and its audience marker on each row

   [Regulation (EU) 2024/1689, Annex XI and Annex XII; GPAI Code of Practice, Transparency chapter, Measure 1.1 and Model Documentation Form; PR-11-DOC step 8]

5. Build one row per Copyright chapter measure and record where the policy document addresses each.

   Record Copyright policy presence rows: one per Copyright chapter measure, with the location in the policy document

   [Regulation (EU) 2024/1689, Art. 53(1)(c); GPAI Code of Practice, Copyright chapter, Measures 1.1 to 1.5]

6. Mark each section of the Commission template present with content, placeholder or absent in the published summary.

   Record Training-content summary structure check: one row per template section

   [Regulation (EU) 2024/1689, Art. 53(1)(d); PR-11-DOC step 39]

7. For a model with systemic risk, add one row per Art. 55(1) point and per Annex XI Section 2 element.

   Record Art. 55 evidence rows: Framework section, Model Report section and PR-15-ROB record per Art. 55(1) point

   [GPAI Code of Practice, Safety and Security chapter, Measure 1.1, Measures 7.1 to 7.3, Measure 9.2 and Measure 10.1; Regulation (EU) 2024/1689, Annex XI Section 2]

Example for steps 4 to 7, a synthetic evidence index for a model named aurora-12b, a placeholder. Each entry names the article point, the Code measure it rests on and the document held with its identifier; the required_when key marks the entries that exist only for a model with systemic risk. The document identifiers, the version numbers and the URL are placeholders.

# synthetic example: run/evidence-index.yaml for aurora-12b (placeholder model name)
model: aurora-12b
model_version: "sha256:synthetic-3b7e"
records:
  provider_determination: PD-01        # PR-CLS output
  open_source_exception: OS-01         # PR-CLS output
  training_compute: TC-01              # PR-CLS output
provider_status: provider
open_source_exception: false
systemic_risk: true
means_of_demonstration:
  route: code-of-practice              # or alternative-adequate-means, Art. 53(4) and Art. 55(2)
  chapters_signed: [transparency, copyright, safety-and-security]
art_53_1:
  a:
    duty: technical documentation with the Annex XI content, kept up to date, given to the AI Office on request
    code_measure: Transparency chapter, Measures 1.1 to 1.3
    documents:
      - id: MDF-03
        kind: Model Documentation Form
        version: 3
        audience: [AIO, NCAs]
        sections:
          - General information
          - Model properties
          - Methods of distribution and licenses
          - Use
          - Training process
          - Information on the data used for training, testing, and validation
          - Computational resources (during training)
          - Energy consumption (during training and inference)
      - id: AXI-S2-01
        kind: Annex XI Section 2 description
        code_measure: Safety and Security chapter, Measure 10.1
        required_when: systemic_risk
  b:
    duty: information and documentation with the Annex XII content for downstream providers
    code_measure: Transparency chapter, Measure 1.2, second paragraph
    documents:
      - id: MDF-03
        kind: Model Documentation Form, fields marked DPs
        audience: [DPs]
  c:
    duty: policy to comply with Union copyright law, identifying rights reservations under Art. 4(3) of Directive (EU) 2019/790
    code_measure: Copyright chapter, Measures 1.1 to 1.5
    documents:
      - id: CP-02
        kind: copyright policy, single document
        sections:
          - Measure 1.1 policy scope and assigned responsibilities
          - Measure 1.2 lawful access and excluded websites
          - Measure 1.3 robots.txt and other machine-readable reservation protocols
          - Measure 1.4 output safeguards and prohibited uses
          - Measure 1.5 point of contact and complaints mechanism
  d:
    duty: public summary of the training content in the AI Office template
    code_measure: none; Commission template published for Art. 53(1)(d)
    documents:
      - id: TCS-01
        kind: public training-content summary
        sections: [General information, List of data sources, Relevant data processing aspects]
        published_at: https://example.invalid/aurora-12b/training-content-summary
art_55_1:
  a:
    duty: model evaluation with state-of-the-art protocols, including documented adversarial testing
    code_measure: Safety and Security chapter, Measure 3.2 and Measure 7.3, point (1)(e) and (f)
    documents:
      - { id: SSF-02, kind: Safety and Security Framework, sections: [Measure 1.1 point (1) trigger points, Measure 1.1 point (2)(a) acceptance criteria] }
      - { id: MR-02, kind: Safety and Security Model Report, sections: [Measure 7.3 point (1)(e) evaluation results, Measure 7.3 point (1)(f) random samples] }
      - { id: MER-01, kind: model evaluation report per systemic risk (PR-15-ROB output) }
      - { id: ATC-01, kind: adversarial testing campaign record (PR-15-ROB output) }
  b:
    duty: assessment and mitigation of systemic risks at Union level, including their sources
    code_measure: Safety and Security chapter, Commitments 2 to 5 and Measure 7.2
    documents:
      - { id: SSF-02, kind: Safety and Security Framework, sections: [Measure 1.1 point (2)(b) mitigations per tier] }
      - { id: MR-02, kind: Safety and Security Model Report, sections: [Measure 7.2 reasons for proceeding, Measure 7.3 point (2) safety mitigations] }
  c:
    duty: serious incident tracking, documentation and reporting to the AI Office
    code_measure: Safety and Security chapter, Measures 9.1 to 9.4
    documents:
      - { id: SIR-01, kind: serious incident register and reporting procedure }
  d:
    duty: adequate cybersecurity for the model and its physical infrastructure
    code_measure: Safety and Security chapter, Measures 6.1 and 6.2 with Appendix 4
    documents:
      - { id: MR-02, kind: Safety and Security Model Report, sections: [Measure 7.3 point (3) Security Goal and security mitigations] }
      - { id: CSD-01, kind: model and infrastructure cybersecurity measures description }

8. Mark each Annex XI and XII row present, partial or absent, with the location of the content.

   Record Annex XI and XII gap table: status, location and record reference per row; the provider's reason for each row marked not applicable

   [Regulation (EU) 2024/1689, Art. 53(1)(a) and Art. 53(1)(b); PR-11-DOC step 11]

9. Run the index check: every Art. 53(1) point holds a document; Art. 55 entries exist only with systemic risk.

   Record Evidence index check output: one line per missing document or misplaced Art. 55 entry

   [Regulation (EU) 2024/1689, Art. 53(1), Art. 53(2) and Art. 55(1); AIAssuranceLabs procedure]

Example for step 9, a check with python and the PyYAML package that reads the index above. It fails when an Art. 53(1) point holds no document, applies the Art. 53(2) exception only to points (a) and (b) and only when the model has no systemic risk, and fails when Art. 55 entries are present for a model whose PR-CLS record does not classify it with systemic risk or absent for one that does. The file name and the exit codes are the example's own.

# synthetic example: check_evidence_index.py, run as
#   python3 check_evidence_index.py run/evidence-index.yaml
import sys

import yaml  # PyYAML

POINTS = ["a", "b", "c", "d"]

with open(sys.argv[1], encoding="utf-8") as f:
    idx = yaml.safe_load(f)

systemic = bool(idx.get("systemic_risk"))
exempt = bool(idx.get("open_source_exception")) and not systemic  # Art. 53(2)
errors = []

art53 = idx.get("art_53_1") or {}
for point in POINTS:
    if point in ("a", "b") and exempt:
        continue  # Art. 53(2) lifts (a) and (b) only; (c) and (d) remain due
    docs = [
        d for d in (art53.get(point) or {}).get("documents") or []
        if d.get("required_when") != "systemic_risk" or systemic
    ]
    if not docs:
        errors.append(f"Art. 53(1)({point}): no document held")
    for d in docs:
        if not d.get("id"):
            errors.append(f"Art. 53(1)({point}): document without an identifier")

art55 = idx.get("art_55_1") or {}
if systemic:
    for point in POINTS:
        if not (art55.get(point) or {}).get("documents"):
            errors.append(f"Art. 55(1)({point}): no document held for a model with systemic risk")
    section_2 = [
        d for d in (art53.get("a") or {}).get("documents") or []
        if d.get("required_when") == "systemic_risk"
    ]
    if not section_2:
        errors.append("Annex XI Section 2: no description held for a model with systemic risk")
elif art55:
    errors.append("Art. 55(1): entries present for a model not classified with systemic risk; check PR-CLS record")

for e in errors:
    print("FAIL", e)
print("evidence index: " + ("complete" if not errors else f"{len(errors)} finding(s)"))
sys.exit(1 if errors else 0)

10. Compare architecture, parameters, modality, format, release date, distribution and licence across Annex XI, Annex XII and the model card.

    Record Cross-document consistency table: Annex XI documentation, Annex XII information, public model card

    [Regulation (EU) 2024/1689, Annex XI Section 1 point 1 and Annex XII point 1; PR-11-DOC step 36]

11. Check that each evaluation result cited in Annex XI or the Model Report resolves to a stored record.

    Record Evaluation result traceability rows: protocol name, dataset version, model version, resolved record identifier

    [Regulation (EU) 2024/1689, Art. 53(1)(a) and Annex XI; GPAI Code of Practice, Safety and Security chapter, Measure 7.3, point (1)(e); PR-11-DOC step 38]

12. Write the completeness and consistency report with the version, owner and update trigger of each indexed document.

    Record Annex XI and XII completeness and consistency report (PR-11-DOC output), with the document control version and update trigger per document

    [GPAI Code of Practice, Transparency chapter, Measures 1.1 and 1.3; PR-11-DOC steps 42 and 44]

A pitfall that vendor documentation omits

The Model Documentation Form covers Annex XI Section 1 and Annex XII, and nothing else. The Transparency chapter of the published GPAI Code of PracticePublished states in the scope paragraph that follows its recitals that the documentation obligations concerning Annex XI Section 2, which apply only to a model with systemic risk, are covered by Measure 10.1 of the Safety and Security chapter; that measure asks for a detailed description of the model architecture, of its integration into AI systems, of the model evaluations conducted with their results and strategies, and of the safety mitigations implemented, retained for at least ten years after the model is placed on the market. An evidence index built from the completed Form alone therefore leaves the Art. 53(1)(a) entry incomplete for every model above the Art. 51 gate, which is why the AXI-S2-01 entry in the example carries required_when: systemic_risk and why the check in step 9 fails on its absence. The Form adds a second trap of the same kind: each field carries a marker for the AI Office, national competent authorities or downstream providers, and several fields exist twice at different precision, such as the amount of computation used for training, recorded to its order of magnitude for national competent authorities and to at least two significant figures for the AI Office, with a measurement methodology field that the Form requires in the absence of a delegated act under Art. 53(5). The Art. 53(1)(b) pack is the subset of fields marked for downstream providers, so a row in the gap table of step 8 is complete only when the content exists at the precision the marker asks for.

The Art. 53(2) exception is narrower than a licence badge suggests. It lifts points (a) and (b) of Art. 53(1) for a model released under a free and open-source licence whose parameters, architecture information and usage information are public, and it falls away for a model with systemic risk; points (c) and (d) remain due in every case. The Commission's notice on the training-content summary template states that the obligation applies to providers of models released under free and open-source licences, that use of the template is mandatory, that the summary is published on the provider's website no later than when the model is placed on the Union market, and that a modifier fills the template for the training content of the modification only; for scraped online content the template asks for the crawlers used, the collection period and a list of the top ten percent of domains scraped, with a smaller share for SMEs. The check in step 9 encodes the exception as written: exempt is true only when the PR-CLS open-source exception record says so and the model has no systemic risk, and it skips points (a) and (b) only. No published document sets a completeness score for any of this; the source hierarchy runs from a harmonised standard cited in the Official Journal, of which none is cited as of the anchor date, through JTC 21 drafts, of which none covers general-purpose AI models, to the code-of-practice level in use here, and the provider's reasons for each row marked not applicable sit in the gap table of step 8.

What the result shows and what it does not

The evidence index and the gap table show presence, cross-document consistency and traceability at the frozen input set; the Limitations section of PR-11-DOC v1.0 states that a change to the documentation after the freeze is outside the result, which step 12 records as an update trigger per document. The protocol does not judge whether the Annex XII information gives a downstream provider a good understanding of the model under Art. 53(1)(b), does not assess the content of the copyright policy under Art. 53(1)(c) beyond the presence of each measure, and does not verify the compute and energy figures under Annex XI Section 1 points 2(d) and 2(e) beyond the records behind them. On the Art. 55 side the index records where the Framework, the Model Report and the PR-15-ROB reports sit; PR-15-ROB v1.0 runs the evaluations the provider's Framework names, is the independent external evaluation of Appendix 3.5 of the Safety and Security chapter only where the engagement is scoped as one, and does not test the physical infrastructure under Art. 55(1)(d); the campaign procedure itself is in the jailbreak-resistance guide and is not repeated here. The index is evidence for the provider's own file, assembled as the Annex XI technical documentation contribution and Annex XII downstream information pack of the GPAI evaluation support service within the limits stated on what we cannot do: no statement of adherence to the Code on the provider's behalf, since the adherence statement under Art. 53(4) is the provider's declaration, and no assessment of alternative adequate means, since that assessment is the Commission's. Whether the Model Report and the Framework satisfy Art. 55(2) remains a matter between the provider, the AI Office and the provider's counsel.

Related pages

The obligations, the two protocols, the related guide and the hub for the role this guide serves are listed below; the evaluation sequence that feeds the Art. 55 rows is on the model evaluation scenario.

Obligations

• Art. 53
• Art. 55

Protocols

• PR-15-ROB v1.0
• PR-11-DOC v1.0

Related guides

• AI red teaming: running a jailbreak-resistance campaign

ICP hubs

• GPAI providers

Every inquiry receives a reply within 24 hours. Scope and availability are discussed per request.

Ask about scopehello@aiassurancelabs.com